Security and Privacy

Security, Data Handling and Trust Boundaries

• The program runs locally and writes all original JSONL record fields, including review, into the output. Avoid using confidential data on public repositories.
• In real mode, model/tokenizer files can be fetched from Hugging Face; network use is not required when weights are cached locally, and mock mode performs no such download.
• No credential vault, HTTP authentication, PII anonymization, encrypted-at-rest storage or audit trail is implemented.
• Use a trusted model source, pin/verify dependencies, review model licensing, and keep .env, tokens, private datasets and model binaries out of Git.
• Treat input reviews as untrusted text; this is a fixed classification prompt but model responses are still untrusted content. No prompt-injection defense is implemented.

flowchart LR
    In[("Local user-supplied JSONL")] --> V["Validate schema / nonempty text"]
    V --> Host["Local Python worker processes"]
    HF[("External model provider or local HF cache")] -. "Optional model download" .-> Host
    Host --> Out[("Local output JSONL - review text retained")]
    Out --> Policy["User-controlled sharing, retention and access"]
    Note["No HTTP authentication, secrets manager or PII removal implemented"] -.-> Host

Loading